Ranking models balance several competing signals to keep a feed engaging over time. Diversity refers to deliberately mixing topics, creators, and formats so that your feed does not become repetitive or overly narrow. Closely related is the novelty versus relevance tradeoff: novel items feel fresh but may miss your interests, while highly relevant items match your known tastes but can grow boring if overused. Platforms try to thread this tradeoff by injecting some unfamiliarity into each session.
Because people respond more to recent posts, social platforms incorporate recency bias into ranking through time decay, a mechanism that gradually lowers the score of older content unless it keeps earning strong engagement. This is why even older winners get eclipsed by fresh competition, and why a viral video from last week may no longer appear at the top of your feed today.
Behind these signals, recommendation systems draw on patterns shared across groups of users. Cohort behavior matters because people with similar patterns help predict what you might enjoy next, letting models learn from groups rather than only from your individual history. Two foundational techniques enable this. Collaborative filtering recommends items liked by people who behave similarly to you, even when the content itself looks unrelated, while content-based recommendation suggests items that resemble content features such as topic, audio, or visuals that you have engaged with. Most platforms combine the two because collaborative filtering uncovers unexpected items while content-based methods fill in when interaction data is sparse, when items are new, or when a viewer behaves unlike anyone else in the cohort.